How can I optimize multi-batch and parallel inference in TensorRT for faster performance on high-resolution image patches?

Description

I am encountering performance bottlenecks while running multi-threaded inference on high-resolution images using TensorRT. The model involves breaking the image into patches to manage GPU memory, performing inference on each patch, and then merging the results. However, the inference time per patch is still high, even when increasing the batch size. Additionally, loading multiple engines onto the GPU to parallelize the inference does not yield the expected speedup. I am seeking advice on optimizing the inference process for faster execution, either by improving batch processing or enabling better parallelism in TensorRT.

Environment

TensorRT Version: 10.5.0
GPU Type: RTX 3050TI 4GB
Nvidia Driver Version: 535.183.01
CUDA Version: 12.2
CUDNN Version: N/A
Operating System + Version: Ubuntu 20.04
Python Version: 3.11
TensorFlow Version: N/A
PyTorch Version: N/A
Baremetal or Container (if container, which image + tag): Baremetal

Relevant Files

`build_engine.py`

def build_engine(onnx_file_path, engine_file_path):
    logger = trt.Logger(trt.Logger.ERROR)
    builder = trt.Builder(logger)
    network = builder.create_network(1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH))
    profile = builder.create_optimization_profile()
    config = builder.create_builder_config()
    parser = trt.OnnxParser(network, logger)

    if not os.path.exists(onnx_file_path):
        print("Failed finding ONNX file!")
        return
    print("Succeeded finding ONNX file!")

    with open(onnx_file_path, 'rb') as model:
        if not parser.parse(model.read()):
            print('Failed parsing the ONNX file')
            for error in range(parser.num_errors):
                print(parser.get_error(error))
            return
    print('Completed parsing of ONNX file')

    # Configure input profile
    input_tensor = network.get_input(0)
    profile.set_shape(input_tensor.name, (min_batch, shape[1], shape[2], shape[3]), shape, (max_batch, shape[1], shape[2], shape[3]))
    config.add_optimization_profile(profile)

    # Build the serialized engine
    engine_string = builder.build_serialized_network(network, config)
    if engine_string is None:
        print("Failed building engine!")
        return
    print("Succeeded building engine!")

    with open(engine_file_path, "wb") as f:
        f.write(engine_string)

`inference.py`

class TRTModel:
    def __init__(self, trt_path):
        self.trt_path = trt_path
        trt.init_libnvinfer_plugins(None, "")
        self.logger = trt.Logger(trt.Logger.ERROR)
        with open(self.trt_path, "rb") as f:
            engine_data = f.read()
        self.engine = trt.Runtime(self.logger).deserialize_cuda_engine(engine_data)

    def create_execution_context(self):
        return self.engine.create_execution_context()

    def process_async(self, input_data):
        _, stream = cudart.cudaStreamCreate()
        context = self.create_execution_context()

        input_size = input_data.nbytes
        output_size = input_data.nbytes

        input_device = cudart.cudaMallocAsync(input_size, stream)[1]
        output_device = cudart.cudaMallocAsync(output_size, stream)[1]

        input_data_np = input_data.cpu().numpy()

        cudart.cudaMemcpyAsync(input_device, input_data_np.ctypes.data, input_data.nbytes,
                               cudart.cudaMemcpyKind.cudaMemcpyHostToDevice, stream)

        context.set_tensor_address('images', int(input_device))
        context.set_tensor_address('output', int(output_device))
        context.execute_async_v3(stream_handle=int(stream))

        output_host = np.empty_like(input_data_np, dtype=np.float32)
        cudart.cudaMemcpyAsync(output_host.ctypes.data, output_device, output_host.nbytes,
                               cudart.cudaMemcpyKind.cudaMemcpyDeviceToHost, stream)
        cudart.cudaStreamSynchronize(stream)

        cudart.cudaFree(input_device)
        cudart.cudaFree(output_device)
        cudart.cudaStreamDestroy(stream)

        return output_host

Steps To Reproduce

Build the Engine: Use build_engine to convert an ONNX model into a TensorRT engine.
Run Inference: Use TRTModel to perform inference on cropped image patches.
Expected Result: While batch sizes are increased, the inference time per patch remains high. Running multiple engines for parallel inference also does not improve performance.
Profiling Results:
- Transfer to device: 0.48 ms
- Inference time: 784.75 ms
- Transfer to host: 0.67 ms
- Total time for a single patch (256x256): 19-22 seconds on average

I am seeking optimization suggestions for improving multi-batch processing or multi-threaded parallel inference in TensorRT.

NVIDIA / TensorRT